This commit limits the `index.translog.sync_interval` to a value not less than `100ms` and
removes the support for fsync on every operation which used to be enabled if `index.translog.sync_interval` was set to `0s`
Now this pr also only schedules an async fsync if the durability is set to `async`. By default not async task is scheduled.
Closes#16152
Today we have two variants of translogs for indexing. We only recommend the buffered
one which also has a 20% advantage in indexing speed. This commit removes the option and defaults
to the buffered case. It also hard-wires the translog buffer to 8kb instead of 64kb. We used to
adjust that buffer based on if the shard is active or not, this code has also been removed and
instead we just keep an 8kb buffer arround.
This commit removes `index.translog.flush_threshold_ops` and `index.translog.disable_flush`
in favor of `index.translog.flush_threshold_size`. The number of operations is meaningless by itself and
can easily be turned into a size value with knowledge of the data. Disabling the flush is only useful in
tests and we can set the size value to a really high value. If users really need to do this they can
also apply a very high value like `1PB`.
- moves calculation of the delay to a single place (ReplicaShardAllocator)
- reduces coupling between GatewayAllocator and RoutingService
- in master failover situations, elapsed delay time is forgotten
Closes#14808
This adds the `cluster.routing.allocation.total_shards_per_node`
setting, which limits the total number of shards across all indices on
each node. It defaults to -1 and can be dynamically configured.
Resolves#14456
This commit moves the size and ops based flush into a synchronous API into
IndexShard and removes the time-based flush alltogether since it' basically
covered by the inactive async flush API we have today. The functionality doesn't
need to be covered by scheduled task and async APIs while we can actually make all
the decisions in a sync manner which is way easier to control and to test.
Closes#13707
Allocation filtering by IP only works today using the node host address. But in some cases, you might want to filter using the publish address which could be different.
Instead of logging the entire `_source` in the indexing slowlog we log by
default just the first 1000 characters - this is controlled by the
`index.indexing.slowlog.source` settings and can be set to `true` to log the
whole `_source`, `false` to log none of it, and a number to log at most that
many characters.
Closes#4485
Change the default delayed allocation timeout from 0 (no delayed allocation) to 1m. The value came from a test of having a node with 50 shards being indexed into (so beefy translog requiring flush on shutdown), then shutting it down and starting it back up and waiting for it to join the cluster. This took, on a slow machine, about 30s.
The value is conservatively low and does not try to address a virtual machine / OS restart for now, in order to not have the affect of node going away and users being concerned that shards are not being allocated to the rest of the cluster as a result of that. The setting can always be changed in order to increase the delayed allocation if needed.
closes#12166
Now that doc values are the default for fielddata, specialized in-memory
formats are becoming an esoteric option. This commit removes such formats:
- `fst` on string fields,
- `compressed` on geo points.
I also removed documentation and tests that the fielddata cache is shared if
you change the format, since this is only true for in-memory fielddata formats
(given that for doc values, the caching is done directly in Lucene).
Today we provide the ability to plug in MergePolicy and
we provide the once lucene ships with. We do not recommend to change
the default and even only a small number of expert users would ever touch
this. This commit removes the ancient log byte size and log doc count
merge policy providers, simplifies the MergePolicy wiring and makes the
tiered MP the one and only default. All notions of a merge policy has been
removed from the docs and should be deprecated in the previous version.
Closes#11588
Using files that must be specified on each node is an anti-pattern
from the API based goal of ES. This change removes the ability
to specify the default mapping with a file on each node.
closes#10620
In order to safely complete recoveries / relocations we have to keep all operation done since the recovery start at available for replay. At the moment we do so by preventing the engine from flushing and thus making sure that the operations are kept in the translog. A side effect of this is that the translog keeps on growing until the recovery is done. This is not a problem as we do need these operations but if the another recovery starts concurrently it may have an unneededly long translog to replay. Also, if we shutdown the engine for some reason at this point (like when a node is restarted) we have to recover a long translog when we come back.
To void this, the translog is changed to be based on multiple files instead of a single one. This allows recoveries to keep hold to the files they need while allowing the engine to flush and do a lucene commit (which will create a new translog files bellow the hood).
Change highlights:
- Refactor Translog file management to allow for multiple files.
- Translog maintains a list of referenced files, both by outstanding recoveries and files containing operations not yet committed to Lucene.
- A new Translog.View concept is introduced, allowing recoveries to get a reference to all currently uncommitted translog files plus all future translog files created until the view is closed. They can use this view to iterate over operations.
- Recovery phase3 is removed. That phase was replaying operations while preventing new writes to the engine. This is unneeded as standard indexing also send all operations from the start of the recovery to the recovering shard. Replay all ops in the view acquired in recovery start is enough to guarantee no operation is lost.
- IndexShard now creates the translog together with the engine. The translog is closed by the engine on close. ShadowIndexShards do not open the translog.
- Moved the ownership of translog fsyncing to the translog it self, changing the responsible setting to `index.translog.sync_interval` (was `index.gateway.local.sync`)
Closes#10624
* Removed the docs for `index.compound_format` and `index.compound_on_flush` - these are expert settings which should probably be removed (see https://github.com/elastic/elasticsearch/issues/10778)
* Removed the docs for `index.index_concurrency` - another expert setting
* Labelled the segments verbose output as experimental
* Marked the `compression`, `precision_threshold` and `rehash` options as experimental in the cardinality and percentile aggs
* Improved the experimental text on `significant_terms`, `execution_hint` in the terms agg, and `terminate_after` param on count and search
* Removed the experimental flag on the `geobounds` agg
* Marked the settings in the `merge` and `store` modules as experimental, rather than the modules themselves
Closes#10782
This commit brings the benefits of the `count` search type to search requests
that have a `size` of 0:
- a single round-trip to shards (no fetch phase)
- ability to use the query cache
Since `count` now provides no benefits over `query_then_fetch`, it has been
deprecated.
Close#7630
We now have a very useful annotation to mark features or parameters as
experimental. Let's use it! This commit replaces some custom text warnings with
this annotation and adds this annotation to some existing features/parameters:
- inner_hits (unreleased yet)
- terminate_after (released in 1.4)
- per-bucket doc count errors in the terms agg (released in 1.4)
I also tagged with this annotation settings which should either be not needed
(like the ability to evict entries from the filter cache based on time) or that
are too deep into the way that Elasticsearch works like the Directory
implementation or merge settings.
Close#9563
This particular note was about fielddata filtering but could cause confusion
that fields that have doc values enabled cannot be used for filtering (as in
a `filtered_query`).
When using the DiskThresholdDecider, it's possible that shards could
already be marked as relocating to the node being evaluated. This commit
adds a new setting `cluster.routing.allocation.disk.include_relocations`
which adds the size of the shards currently being relocated to this node
to the node's used disk space.
This new option defaults to `true`, however it's possible to
over-estimate the usage for a node if the relocation is already
partially complete, for instance:
A node with a 10gb shard that's 45% of the way through a relocation
would add 10gb + (.45 * 10) = 14.5gb to the node's disk usage before
examining the watermarks to see if a new shard can be allocated.
Fixes#7753
Relates to #6168
This documentation was dangerous because it felt like it was possible to gain
substantial performance by just switching the codec of the index.
However, non-default codecs are dangerous to use since they are not supported
in terms of backward compatibility, and most improvements that they bring have
been folded into the default codec anyway (for example, the default codec
"pulses" postings lists that contain a single document).
Adds a breaker for request BigArrays, which are used for parent/child
queries as well as some aggregations. Certain operations like Netty HTTP
responses and transport responses increment the breaker, but will not
trip.
This also changes the output of the nodes' stats endpoint to show the
parent breaker as well as the fielddata and request breakers.
There are a number of new settings for breakers now:
`indices.breaker.total.limit`: starting limit for all memory-use breaker,
defaults to 70%
`indices.breaker.fielddata.limit`: starting limit for fielddata breaker,
defaults to 60%
`indices.breaker.fielddata.overhead`: overhead for fielddata breaker
estimations, defaults to 1.03
(the fielddata breaker settings also use the backwards-compatible
setting `indices.fielddata.breaker.limit` and
`indices.fielddata.breaker.overhead`)
`indices.breaker.request.limit`: starting limit for request breaker,
defaults to 40%
`indices.breaker.request.overhead`: request breaker estimation overhead,
defaults to 1.0
The breaker service infrastructure is now generic and opens the path to
adding additional circuit breakers in the future.
Fixes#6129
Conflicts:
src/main/java/org/elasticsearch/index/fielddata/IndexFieldData.java
src/main/java/org/elasticsearch/index/fielddata/IndexFieldDataService.java
src/main/java/org/elasticsearch/index/fielddata/RamAccountingTermsEnum.java
src/main/java/org/elasticsearch/index/fielddata/ordinals/GlobalOrdinalsBuilder.java
src/main/java/org/elasticsearch/index/fielddata/ordinals/InternalGlobalOrdinalsBuilder.java
src/main/java/org/elasticsearch/index/fielddata/plain/AbstractIndexOrdinalsFieldData.java
src/main/java/org/elasticsearch/index/fielddata/plain/DisabledIndexFieldData.java
src/main/java/org/elasticsearch/index/fielddata/plain/IndexIndexFieldData.java
src/main/java/org/elasticsearch/index/fielddata/plain/NonEstimatingEstimator.java
src/main/java/org/elasticsearch/index/fielddata/plain/PackedArrayIndexFieldData.java
src/main/java/org/elasticsearch/index/fielddata/plain/ParentChildIndexFieldData.java
src/main/java/org/elasticsearch/index/fielddata/plain/SortedSetDVOrdinalsIndexFieldData.java
src/main/java/org/elasticsearch/node/internal/InternalNode.java
src/test/java/org/elasticsearch/index/aliases/IndexAliasesServiceTests.java
src/test/java/org/elasticsearch/index/codec/CodecTests.java
src/test/java/org/elasticsearch/index/fielddata/AbstractFieldDataTests.java
src/test/java/org/elasticsearch/index/fielddata/IndexFieldDataServiceTests.java
src/test/java/org/elasticsearch/index/mapper/MapperTestUtils.java
src/test/java/org/elasticsearch/index/query/IndexQueryParserFilterCachingTests.java
src/test/java/org/elasticsearch/index/query/SimpleIndexQueryParserTests.java
src/test/java/org/elasticsearch/index/query/guice/IndexQueryParserModuleTests.java
src/test/java/org/elasticsearch/index/search/FieldDataTermsFilterTests.java
src/test/java/org/elasticsearch/index/search/child/ChildrenConstantScoreQueryTests.java
src/test/java/org/elasticsearch/index/similarity/SimilarityTests.java
This change just changes the default for index.codec.bloom.load to
false: with recent performance improvements to ID lookup, such as
#6298, bloom filters don't give much of a performance gain anymore,
and they can consume non-trivial RAM when there are many tiny
documents.
For now, we still index the bloom filters, so if a given app wants
them back, it can just update the index.codec.bloom.load to true.
Closes#6959
`mmapfs` is really good for random access but can have sideeffects if
memory maps are large depending on the operating system etc. A hybrid
solution where only selected files are actually memory mapped but others
mostly consumed sequentially brings the best of both worlds and
minimizes the memory map impact.
This commit mmaps only the `dvd` and `tim` file for fast random access
on docvalues and term dictionaries.
Closes#6636
This commit upgrades to the latest Lucene 4.8.1 release including the
following bugfixes:
* An IndexThrottle now kicks in when merges start falling behind
limiting index threads to 1 until merges caught up. Closes#6066
* RateLimiter now kicks in at the configured rate where previously
the limiter was limiting at ~8MB/sec almost all the time. Closes#6018
It's dangerous to expose SerialMergeScheduler as an option: since it only allows one merge at a time, it can easily cause merging to fall behind.
Closes#6120
The current setting of 20MB/sec seems to be too conservative given
the capabilities of modern hardware. Even on cloud infrastructure this
seems to be too lowish. A 50MB default should provide better out of the box
performance
Currently we use 5k operations as a flush threshold. Indexing 5k documents
per second is rather common which would cause the index to be committed on
the lucene level each time the flush logic runs which is 5 seconds by default.
We should rather use a size based threshold similar to the lucene index writer
that doesn't cause such agressive commits which can slow down indexing significantly
especially since they cause the underlying devices to fsync their data.
Load tests showed that SerialMS has problems to keep up with
the merges under high load. We should switch back to CMS
until we have a better story to balance merge
threads / efforts across shards on a single node.
Closes#5817
Today, we use ConcurrentMergeScheduler, and this can be painful since it is concurrent on a shard level, with a max of 3 threads doing concurrent merges. If there are several shards being indexed, then there will be a minor explosion of threads trying to do merges, all being throttled by our merge throttling.
Moving to serial merge scheduler will still maintain concurrency of merges across shards, as we have the merge thread pool that schedules those merges. It will just be a serial one on a specific shard.
Also, on serial merge scheduler, we now have a limit of how many merges it will do at one go, so it will let other shards get their fair chance of merging. We use the pending merges on IW to check if merges are needed or not for it.
Note, that if a merge is happening, it will not block due to a sync on the maybeMerge call at indexing (flush) time, since we wrap our merge scheduler with the EnabledMergeScheduler, where maybeMerge is not activated during indexing, only with explicit calls to IW#maybeMerge (see Merges).
closes#5447
Removed unused misc.asciidoc file
Added plugins directory to directory layout
Fixed transport.tcp.connect_timeout value to match the code found in NetworkService.TcpSettings
Clarified that phrase query does not preserve order of terms
Clarified merge page
Added instructions on how to build documentation to docs/README
* Clean up s/ElasticSearch/Elasticsearch on docs/*
* Clean up s/ElasticSearch/Elasticsearch on src/* bin/* & pom.xml
* Clean up s/ElasticSearch/Elasticsearch on NOTICE.txt and README.textile
Closes#4634
This adds the field data circuit breaker, which is used to estimate
the amount of memory required to load field data before loading it. It
then raises a CircuitBreakingException if the limit is exceeded.
It is configured with two parameters:
`indices.fielddata.cache.breaker.limit` - the maximum number of bytes
of field data to be loaded before circuit breaking. Defaults to
`indices.fielddata.cache.size` if set, unbounded otherwise.
`indices.fielddata.cache.breaker.overhead` - a contast for all field
data estimations to be multiplied with before aggregation. Defaults to
1.03.
Both settings can be configured dynamically using the cluster update
settings API.
This commits add doc values support to geo point using the exact same approach
as for numeric data: geo points for a given document are stored uncompressed
and sequentially in a single binary doc values field.
Close#4207
This commit changes field data configuration updates so that they are
immediately taken into account for loading new segments. The way it works
is that field data configuration is now cached separately from the field
data cache, meaning that it is now possible to clear the field data
configuration from IndexFieldDataService while the cache will stay around. On
the next time that Elasticsearch will reload field data configuration, it will
check if there is already a cache entry, and reuse it if it exists.
To disable field data loading, all that is required is to change the field
data format to "none" (supported by all field data types) using the update
mapping API. Elasticsearch will then refuse to load field data on any new
segment, but field data which has been loaded on the previous segments will
remain available. So you need to clear the field data cache in order to
reclaim memory (otherwise memory will be reclaimed slower, as segments get
merged).
Close#4430Close#4431
This commit allows for using Lucene doc values as a backend for field data,
moving the cost of building field data from the refresh operation to indexing.
In addition, Lucene doc values can be stored on disk (partially, or even
entirely), so that memory management is done at the operating system level
(file-system cache) instead of the JVM, avoiding long pauses during major
collections due to large heaps.
So far doc values are supported on numeric types and non-analyzed strings
(index:no or index:not_analyzed). Under the hood, it uses SORTED_SET doc values
which is the only type to support multi-valued fields. Since the field data API
set is a bit wider than the doc values API set, some operations are not
supported:
- field data filtering: this will fail if doc values are enabled,
- field data cache clearing, even for memory-based doc values formats,
- getting the memory usage for a specific field,
- knowing whether a field is actually multi-valued.
This commit also allows for configuring doc-values formats on a per-field basis
similarly to postings formats. In particular the doc values format of the
_version field can be configured through its own field mapper (it used to be
handled in UidFieldMapper previously).
Closes#3806
* Merged segments are now warmed-up at the end of the merge operation instead
of _refresh, so that _refresh doesn't pay the price for the warm-up of merged
segments, which is often higher than flushed segments because of their size.
* Even when no _warmer is registered, some basic warm-up of the segments is
performed: norms, doc values (_version). This should help a bit people who
forget to register warmers.
* Eager loading support for the parent id cache and field data: when one
can't predict what terms will be present in the index, it is tempting to use
a match_all query in a warmer, but in that case, query execution might not be
much faster than field data loading so having a warmer that only loads field
data without running a query can be useful.
Closes#3819
This commit adds two main pieces, the first is a ClusterInfoService
that provides a service running on the master nodes that fetches the
total/free bytes for each data node in the cluster as well as the
sizes of all shards in the cluster. This information is gathered by
default every 30 seconds, and can be changed dynamically by setting
the `cluster.info.update.interval` setting. This ClusterInfoService
can hopefully be used in the future to weight nodes for allocation
based on their disk usage, if desired.
The second main piece is the DiskThresholdDecider, which can disallow
a shard from being allocated to a node, or from remaining on the node
depending on configuration parameters. There are three main
configuration parameters for the DiskThresholdDecider:
`cluster.routing.allocation.disk.threshold_enabled` controls whether
the decider is enabled. It defaults to false (disabled). Note that the
decider is also disabled for clusters with only a single data node.
`cluster.routing.allocation.disk.watermark.low` controls the low
watermark for disk usage. It defaults to 0.70, meaning ES will not
allocate new shards to nodes once they have more than 70% disk
used. It can also be set to an absolute byte value (like 500mb) to
prevent ES from allocating shards if less than the configured amount
of space is available.
`cluster.routing.allocation.disk.watermark.high` controls the high
watermark. It defaults to 0.85, meaning ES will attempt to relocate
shards to another node if the node disk usage rises above 85%. It can
also be set to an absolute byte value (similar to the low watermark)
to relocate shards once less than the configured amount of space is
available on the node.
Closes#3480
Boaz Leskes
Dongjoon Hyun
Robert Muir
Simon Willnauer
Jim Ferenczi
Clinton Gormley
William
Yannick Welsch
Lee Hinman
Jason O'Donnell
David Pilato
Michael McCandless
Nik Everett
Shay Banon
Adrien Grand
Martijn van Groningen
Pascal Borreli
Ryan Ernst
Joshua Rich
Masaru Hasegawa
Itamar Syn-Hershko
Israel Tsadok
Henrik Nordvik
Suyog Rao
Tanvir Alam
Alexander Reelsen
Peter Johnson @insertcoffee
Ian Babrou
Ivan Brusic
Christoph Frick
Kouhei Sutou
Kevin Wang
Konrad Feldmeier
Oleg Anashkin